You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
OS reauthentication, locking and independent agent code review exist, but Chromium-equivalent security has not been established. This work starts alongside backlog item 1 and is a release gate, not the final task in the feature sequence.
Acceptance criteria
Validate the packaged Windows authentication helper and actual macOS Touch ID/Keychain behavior, including upgrades and helper integrity.
Agree the same-user-process threat model, including shell-capable agents. Evaluate stronger Windows key protection if needed; OS reauthentication alone does not change DPAPI protection.
Obtain external security review and track/resolve findings. Clearly distinguish it from the completed independent agent review.
Record evidence by platform and supported build; unsupported configurations must remain locked.
Evidence
See packages/cm-browser/password-security.md and password-review.md. Actual Windows Hello verification is recorded there; do not infer unperformed macOS or packaged validation. No Chromium-equivalence claim until supported by evidence.
Local source audit checkpoint, not external assurance: the ordinary installed Conf/atomically writer already uses a sibling temporary file, fsync and rename, but Conf 15.1.0 falls back to direct in-place overwrite on EXDEV and under SNAP. Because all cm-browser settings/history writes rewrite the same store containing the vault, strict fail-closed writes must cover the whole cm-browser store, not only writeLogins. Next local assurance slice: prevent non-atomic fallback and add synthetic filesystem fault/interruption/reopen tests, retaining complete old-or-new state and migration records. Do not add a second database/journal or silently reset malformed data.
Other local gaps: test real vault-auth adapter branches at OS/child-process boundaries (missing helper, failed/spawn-error/timeout Windows helper; unavailable/rejected Touch ID; unsupported OS), debug-pipe/encryption-unavailable denial and runtime capability loss with ticket invalidation. Existing fixtures replace verify wholesale, so they do not prove these adapter branches. Whole-file malformed JSON refusal is distinct from corrupt ciphertext; preserve unrelated browser data rather than auto-resetting it.
Packaged Windows helper inclusion/architecture can be asserted locally, but helper integrity, actual installed upgrades, native cancellation, real macOS Touch ID/Keychain, signed-build behavior, agreed same-user-process threat model and external security review still need separate evidence. Documented development Windows Hello verification must not be generalized to all builds. No release or installed-process restart authorized by this work. Findings are source-audit results; tests/fixes for this slice have not run yet.
Strict-storage evidence slice published: 80d7a6b (fix(browser): preserve store on atomic write failures), based on ef7cb7f. One commit and one normal push with hooks; refs/heads/browser-tabs verified equal to local HEAD. Commit: 80d7a6b
The entire cm-browser store now uses exclusive sibling temp creation, complete write, file fsync, close and rename, with no EXDEV/SNAP in-place fallback. Existing filename/JSON/literal-key API and unknown fields during key mutations are preserved; other stores stay on electron-store. Malformed existing stores refuse overwrite, and reads are not a success cache. Production does not monkeypatch fs or depend on private Conf methods.
Evidence from packages/desktop:
bun scripts/browser-smoke.ts --persistence-exdev: RED against Conf 15.1.0, then GREEN with strict storage. EXDEV/SNAP reach rename, directWrites=0, rejected=true, oldCompleteStoreSurvives=true. Includes normal SNAP success, partial-write/ENOSPC, fsync/EIO, JSON semantics and malformed-store refusal.
bun scripts/browser-smoke.ts --persistence-reopen: PASS eleven serial disposable Electron phases. Owned child handles are killed immediately before/after rename without graceful cleanup; fresh processes compare complete old/new store bytes and decrypt actual encrypted vaults/settings. Close-open, close-closed and exclusive-open collision preserve original/unowned state. Actual readLogins migration failure/reopen preserves legacy/settings; success/reopen authenticates the new vault with no legacy key.
bun scripts/browser-smoke.ts --accounts: PASS.
bun test src/main/browser: 31 pass, 0 fail, 911 expectations.
bun typecheck: PASS; rerun after the final fixture edit also PASS.
Full bun scripts/browser-smoke.ts: PASS once at the end, serially; expected /fail ERR_EMPTY_RESPONSE and Chromium Mojo diagnostic.
Scoped Prettier and staged git diff --check: PASS, with Git LF/CRLF notices. Scoped oxlint before the final trivial fixture fix: 56 warnings, 0 errors, not warning-free.
Enabled pre-push hook: 31/31 typecheck tasks successful, 29 cached. Completed feature suites were not gratuitously rerun for publication.
Full scoped source review ses_f5905c6a3ffeYlhAvgmZBvPSmg found no production issues; its sole low finding was checkpoint JSON creation/write visibility. Parent fixed only the test fixture: checkpoint.json.tmp is written with wx, then published through the captured rename. Parent reports all eleven reopen phases PASS afterward (owned killed PIDs 12384 and 17204) and Prettier PASS. This is agent source review, not external security assurance.
Limits remain explicit in packages/cm-browser/password-security.md:
Evidence assumes an initialized OS-encryption profile. Immediate app.exit during initial fixture setup caused cross-process decryption failure despite identical store bytes; normal setup app.quit resolved it. First-ever key initialization interruption remains unverified.
Process interruption is not power-loss durability. Parent-directory fsync/durability, hardware/keychain recovery guarantees and multiprocess read-modify-write coordination are not provided.
An interrupted write can leave cm-browser..tmp with encrypted vault material and ordinary settings/legacy metadata. No automatic promotion, broad deletion or secure erasure; ambiguous close may retain a descriptor until process exit.
Authentication/dialog choices are stubbed only in disposable children; real Windows encryption/decryption is exercised. Actual native CredUI/OS-adapter failure and lock termination evidence, keyboard/screen-reader usability, packaged integrity/upgrades, macOS, agreed same-user threat model and external review remain open.
Keeping #34 OPEN: this publishes only the strict-storage slice, not final password-manager assurance or Chromium parity. Installed app/server/bridge untouched; no release or deployment.
Local verification checkpoint, not publication or issue closure. Based on browser-tabs at 4536d8d in the separate cm-browser-review-20260918 checkout; six changed/new files remain uncommitted and unpushed.
Fixed a reproduced capability-loss defect: observing unavailable encryption or remote-debugging port/pipe now revokes pending authentication and existing access tickets. Recovery does not restore authority; OS authentication checks capability before and after completion. Profile publication samples status after capability and avoids recursive locking. The new accounts regression failed before the fix (unlocking instead of locked); completion-time denial separately failed before its guard.
Verification: focused disposable native --accounts regression PASS for all three denial modes; full desktop bun test 505 PASS / 1 SKIP / 0 FAIL, including an isolated child running 10 production authentication-adapter boundary cases; desktop bun typecheck PASS; full disposable browser smoke PASS on Chromium 152.0.7977.78. Native smoke emitted Chromium diagnostic messages despite successful completion. Scoped lint: 67 warnings / 0 errors. Formatting applied and git diff --check passed. Independent Standards and Spec source reviews found no actionable findings.
This covers automated adapter failure handling and observed capability-loss revocation, NOT real CredUI/Touch ID, packaged-helper integrity, physical/platform/accessibility validation, power-loss/key-recovery assurance or external security review. Those gates remain open. Tests used disposable profiles and stubbed OS ceremonies, not real credentials. No installed app/server/bridge restarted; no release, deployment, commit or push.
Published the capability-loss fix as d282cc2 (fix(browser): revoke vault access on capability loss). Six scoped files, separate from #25. Local and remote browser-tabs revisions match; worktree clean.
Unavailable secure storage or remote-debugging port/pipe now revokes pending authentication and existing access tickets. Capability recovery does not restore authority. Authentication checks capability before and after the OS adapter, and profile publication avoids recursive locking.
Publication verification: production authentication-adapter regression passed, including its isolated 10-case child suite; disposable native --accounts passed encryption, remote-debugging-port and remote-debugging-pipe cases. Desktop typecheck passed. The earlier full combined desktop/native results remain recorded in #25 and the prior checkpoint; only focused suites were rerun for publication. Normal push hooks passed 31/31 typecheck tasks, none cached.
#34 remains OPEN. Required evidence still includes packaged Windows helper integrity/upgrades, real macOS Touch ID/Keychain, physical OS-authentication/usability checks, key-loss/corruption/migration/interrupted-write coverage beyond current synthetic evidence, the agreed same-user/shell-capable-agent threat model, and external security review. Automated adapter/source review is not external assurance or power-loss/key-recovery proof. No packaged release, deployment or installed-process restart occurred.
The push also reported 43 Dependabot alerts on the default branch (1 critical, 7 high, 25 moderate, 10 low). These have not been triaged by this slice and are not resolved by the vault fix: https://github.com/trmosala/opencode/security/dependabot
macOS is now the sole target. Accounts, capability loss, atomic-write failure, process-crash/reopen and migration fixtures pass; the new macOS encrypted-backup subprocess fixture passes. Authentication is stubbed in these fixtures. This remains OPEN: actual Touch ID/Keychain ceremonies, packaged helper integrity/upgrades, same-user-process threat acceptance and external security review are outstanding. Installer CI run 35688756323 was blocked before execution by GitHub billing/spending limits. Agent source review is not external assurance. The detailed record also identifies physical entry/accessibility/translation and OS-key-loss limitations.
The local unsigned macOS arm64 build at cc308d72f3 resolves the packaged-helper inclusion/integrity and installed-upgrade evidence slice.
Verified against the installed CookieMonster 1.18.27 PKG:
/Applications/CookieMonster.app/Contents/Resources/vault-auth/macos-entry-arm64 exists, is executable, and is an arm64 Mach-O binary.
The installed helper is byte-identical to packages/desktop/resources/vault-auth/macos-entry-arm64; both have SHA-256 4f51afc0b4f7c1be6f8ff8d817214435f45597ebf05c36ff5ba491bc13966a47.
The PKG completed a real install over the previous build, a second idempotent upgrade, owned CLI removal, full app removal staging, and final reinstall. The final package receipt is com.ogilvy.cookiemonster version 1.18.27.
The installed Electron 44.3.0 app started successfully; bridge health and model discovery passed. The public/private CLI copies were byte-identical and /usr/local/bin/opencode resolved correctly.
Packaging configuration, macOS link ownership, DMG verification, ZIP integrity, and desktop typecheck passed. Gatekeeper rejection is expected because signing and notarization were intentionally disabled for this test build.
This does not close #34. Actual macOS Touch ID success/cancel/failure and Keychain behavior still require physical interaction, and same-user-process threat acceptance plus external security review remain outstanding. A Computer Use attempt to begin physical validation was blocked because the Mac was locked; no biometric, Keychain, accessibility, or external-review claim is made here.
Published the final local macOS assurance follow-up as 7823bcda76.
Passed on this Mac
bun scripts/browser-smoke.ts --persistence-reopen: all 11 real Electron child phases passed. Fresh processes decrypted the encrypted vault, failed/successful migrations preserved the expected complete state, and owned children killed before/after rename reopened to complete old/new stores.
bun scripts/browser-smoke.ts --accounts: encryption loss, remote-debugging port, and remote-debugging pipe revoked access and remained locked.
bun scripts/browser-smoke.ts --persistence-exdev: EXDEV/SNAP refusal, partial-write, fsync-failure, malformed-store, literal-key, and complete-store preservation passed.
Focused production auth/access/backup tests: 9 passed, 0 failed. This covers unavailable/rejected Touch ID, unsupported OS, missing/failing/timed-out helper boundaries, lock invalidation, corruption/tamper refusal, wrong passphrases, and interrupted publication.
Actual login Keychain metadata contains the Electron Safe Storage item used by the isolated Electron harness. No secret was requested or exposed.
Source, packaged, and installed macos-entry-arm64 helpers are executable 54,256-byte arm64 Mach-O files with identical SHA-256 4f51afc0b4f7c1be6f8ff8d817214435f45597ebf05c36ff5ba491bc13966a47.
Installed receipt: com.ogilvy.cookiemonster 1.18.27. The local test build remains intentionally unsigned/unnotarized; helper signature is linker-generated ad hoc.
The security document now states the concrete macOS internal-testing boundary: a same-user shell/debugger/Accessibility/Keychain-capable process, shell-capable agent, compromised Electron main process, or equivalent malware is inside the trusted local-user boundary and can bypass application permission/lock controls. Touch ID is an application gate, not a separately hardware-bound vault key. No Chromium-equivalence, malware-resistance, independent-audit, or external-distribution claim is made.
Physical Touch ID result: BLOCKED
The production adapter reached the real systemPreferences.promptTouchID call on this Mac. The isolated test window was frontmost and now correctly says “Verify with Touch ID.” Two 150-second attempts waited for a physical fingerprint and expired without a result. They are not recorded as success or cancellation passes. The test never opened or decrypted the password vault.
#34 remains open because the user must physically complete success and cancellation, and an external security review with tracked findings has not been supplied. First-key initialization interruption, power-loss/key-loss recovery, physical accessibility, and reviewed translations also remain outside the proven scope. Agent/source review is not being relabeled as external assurance.
This completes the actual Touch ID success/cancel portion alongside the previously recorded Keychain-backed cross-process persistence, capability-loss/corruption/migration/storage-fault, and packaged-helper integrity evidence.
#34 remains open only for assurance that cannot be truthfully supplied by this implementation session: explicit owner/security acceptance of the documented same-user shell-capable-agent boundary, external security review with tracked findings, and the separately documented first-key/power-loss/key-loss recovery limitations. No agent review is being relabeled as external review.
The repository owner explicitly accepted the documented assurance limits for the internal macOS test build on 2026-09-22. The decision is recorded in 7f18709634.
Accepted internal-test scope:
same-user processes with shell, debugger, Accessibility, Keychain, or equivalent authority—including shell-capable agents—are inside the trusted local-user boundary;
external security review is waived for this internal test-only build;
no Chromium-equivalence, malware-resistance, independent-audit, or external-distribution claim is made.
Broader distribution still requires the repository's written WPP Open owner sign-off and external security review with tracked findings.
Together with the previously recorded production-adapter Touch ID VERIFIED/CANCELLED results, actual Keychain-backed cross-process persistence, corruption/migration/interruption/fault coverage, and packaged-helper integrity/upgrade evidence, this completes the owner-approved #34 scope.
Activity
trmosala commentedon Sep 16, 2026
Local source audit checkpoint, not external assurance: the ordinary installed Conf/atomically writer already uses a sibling temporary file, fsync and rename, but Conf 15.1.0 falls back to direct in-place overwrite on EXDEV and under SNAP. Because all cm-browser settings/history writes rewrite the same store containing the vault, strict fail-closed writes must cover the whole cm-browser store, not only writeLogins. Next local assurance slice: prevent non-atomic fallback and add synthetic filesystem fault/interruption/reopen tests, retaining complete old-or-new state and migration records. Do not add a second database/journal or silently reset malformed data.
Other local gaps: test real vault-auth adapter branches at OS/child-process boundaries (missing helper, failed/spawn-error/timeout Windows helper; unavailable/rejected Touch ID; unsupported OS), debug-pipe/encryption-unavailable denial and runtime capability loss with ticket invalidation. Existing fixtures replace verify wholesale, so they do not prove these adapter branches. Whole-file malformed JSON refusal is distinct from corrupt ciphertext; preserve unrelated browser data rather than auto-resetting it.
Packaged Windows helper inclusion/architecture can be asserted locally, but helper integrity, actual installed upgrades, native cancellation, real macOS Touch ID/Keychain, signed-build behavior, agreed same-user-process threat model and external security review still need separate evidence. Documented development Windows Hello verification must not be generalized to all builds. No release or installed-process restart authorized by this work. Findings are source-audit results; tests/fixes for this slice have not run yet.
trmosala commentedon Sep 16, 2026
Strict-storage evidence slice published: 80d7a6b (fix(browser): preserve store on atomic write failures), based on ef7cb7f. One commit and one normal push with hooks; refs/heads/browser-tabs verified equal to local HEAD. Commit: 80d7a6b
The entire cm-browser store now uses exclusive sibling temp creation, complete write, file fsync, close and rename, with no EXDEV/SNAP in-place fallback. Existing filename/JSON/literal-key API and unknown fields during key mutations are preserved; other stores stay on electron-store. Malformed existing stores refuse overwrite, and reads are not a success cache. Production does not monkeypatch fs or depend on private Conf methods.
Evidence from packages/desktop:
Full scoped source review ses_f5905c6a3ffeYlhAvgmZBvPSmg found no production issues; its sole low finding was checkpoint JSON creation/write visibility. Parent fixed only the test fixture: checkpoint.json.tmp is written with wx, then published through the captured rename. Parent reports all eleven reopen phases PASS afterward (owned killed PIDs 12384 and 17204) and Prettier PASS. This is agent source review, not external security assurance.
Limits remain explicit in packages/cm-browser/password-security.md:
Keeping #34 OPEN: this publishes only the strict-storage slice, not final password-manager assurance or Chromium parity. Installed app/server/bridge untouched; no release or deployment.
trmosala commentedon Sep 18, 2026
Local verification checkpoint, not publication or issue closure. Based on browser-tabs at 4536d8d in the separate cm-browser-review-20260918 checkout; six changed/new files remain uncommitted and unpushed.
Fixed a reproduced capability-loss defect: observing unavailable encryption or remote-debugging port/pipe now revokes pending authentication and existing access tickets. Recovery does not restore authority; OS authentication checks capability before and after completion. Profile publication samples status after capability and avoids recursive locking. The new accounts regression failed before the fix (unlocking instead of locked); completion-time denial separately failed before its guard.
Verification: focused disposable native --accounts regression PASS for all three denial modes; full desktop bun test 505 PASS / 1 SKIP / 0 FAIL, including an isolated child running 10 production authentication-adapter boundary cases; desktop bun typecheck PASS; full disposable browser smoke PASS on Chromium 152.0.7977.78. Native smoke emitted Chromium diagnostic messages despite successful completion. Scoped lint: 67 warnings / 0 errors. Formatting applied and git diff --check passed. Independent Standards and Spec source reviews found no actionable findings.
This covers automated adapter failure handling and observed capability-loss revocation, NOT real CredUI/Touch ID, packaged-helper integrity, physical/platform/accessibility validation, power-loss/key-recovery assurance or external security review. Those gates remain open. Tests used disposable profiles and stubbed OS ceremonies, not real credentials. No installed app/server/bridge restarted; no release, deployment, commit or push.
fix(browser): revoke vault access on capability loss
trmosala commentedon Sep 19, 2026
Published the capability-loss fix as d282cc2 (fix(browser): revoke vault access on capability loss). Six scoped files, separate from #25. Local and remote browser-tabs revisions match; worktree clean.
Unavailable secure storage or remote-debugging port/pipe now revokes pending authentication and existing access tickets. Capability recovery does not restore authority. Authentication checks capability before and after the OS adapter, and profile publication avoids recursive locking.
Publication verification: production authentication-adapter regression passed, including its isolated 10-case child suite; disposable native --accounts passed encryption, remote-debugging-port and remote-debugging-pipe cases. Desktop typecheck passed. The earlier full combined desktop/native results remain recorded in #25 and the prior checkpoint; only focused suites were rerun for publication. Normal push hooks passed 31/31 typecheck tasks, none cached.
#34 remains OPEN. Required evidence still includes packaged Windows helper integrity/upgrades, real macOS Touch ID/Keychain, physical OS-authentication/usability checks, key-loss/corruption/migration/interrupted-write coverage beyond current synthetic evidence, the agreed same-user/shell-capable-agent threat model, and external security review. Automated adapter/source review is not external assurance or power-loss/key-recovery proof. No packaged release, deployment or installed-process restart occurred.
The push also reported 43 Dependabot alerts on the default branch (1 critical, 7 high, 25 moderate, 10 low). These have not been triaged by this slice and are not resolved by the vault fix: https://github.com/trmosala/opencode/security/dependabot
trmosala commentedon Sep 22, 2026
Validated on macOS arm64 in commit cc308d7. Detailed evidence and limitations.
macOS is now the sole target. Accounts, capability loss, atomic-write failure, process-crash/reopen and migration fixtures pass; the new macOS encrypted-backup subprocess fixture passes. Authentication is stubbed in these fixtures. This remains OPEN: actual Touch ID/Keychain ceremonies, packaged helper integrity/upgrades, same-user-process threat acceptance and external security review are outstanding. Installer CI run 35688756323 was blocked before execution by GitHub billing/spending limits. Agent source review is not external assurance. The detailed record also identifies physical entry/accessibility/translation and OS-key-loss limitations.
trmosala commentedon Sep 22, 2026
The local unsigned macOS arm64 build at
cc308d72f3resolves the packaged-helper inclusion/integrity and installed-upgrade evidence slice.Verified against the installed CookieMonster 1.18.27 PKG:
/Applications/CookieMonster.app/Contents/Resources/vault-auth/macos-entry-arm64exists, is executable, and is an arm64 Mach-O binary.packages/desktop/resources/vault-auth/macos-entry-arm64; both have SHA-2564f51afc0b4f7c1be6f8ff8d817214435f45597ebf05c36ff5ba491bc13966a47.com.ogilvy.cookiemonsterversion 1.18.27./usr/local/bin/opencoderesolved correctly.This does not close #34. Actual macOS Touch ID success/cancel/failure and Keychain behavior still require physical interaction, and same-user-process threat acceptance plus external security review remain outstanding. A Computer Use attempt to begin physical validation was blocked because the Mac was locked; no biometric, Keychain, accessibility, or external-review claim is made here.
trmosala commentedon Sep 22, 2026
Published the final local macOS assurance follow-up as
7823bcda76.Passed on this Mac
bun scripts/browser-smoke.ts --persistence-reopen: all 11 real Electron child phases passed. Fresh processes decrypted the encrypted vault, failed/successful migrations preserved the expected complete state, and owned children killed before/after rename reopened to complete old/new stores.bun scripts/browser-smoke.ts --accounts: encryption loss, remote-debugging port, and remote-debugging pipe revoked access and remained locked.bun scripts/browser-smoke.ts --persistence-exdev: EXDEV/SNAP refusal, partial-write, fsync-failure, malformed-store, literal-key, and complete-store preservation passed.Electron Safe Storageitem used by the isolated Electron harness. No secret was requested or exposed.macos-entry-arm64helpers are executable 54,256-byte arm64 Mach-O files with identical SHA-2564f51afc0b4f7c1be6f8ff8d817214435f45597ebf05c36ff5ba491bc13966a47.com.ogilvy.cookiemonster1.18.27. The local test build remains intentionally unsigned/unnotarized; helper signature is linker-generated ad hoc.The security document now states the concrete macOS internal-testing boundary: a same-user shell/debugger/Accessibility/Keychain-capable process, shell-capable agent, compromised Electron main process, or equivalent malware is inside the trusted local-user boundary and can bypass application permission/lock controls. Touch ID is an application gate, not a separately hardware-bound vault key. No Chromium-equivalence, malware-resistance, independent-audit, or external-distribution claim is made.
Physical Touch ID result: BLOCKED
The production adapter reached the real
systemPreferences.promptTouchIDcall on this Mac. The isolated test window was frontmost and now correctly says “Verify with Touch ID.” Two 150-second attempts waited for a physical fingerprint and expired without a result. They are not recorded as success or cancellation passes. The test never opened or decrypted the password vault.#34 remains open because the user must physically complete success and cancellation, and an external security review with tracked findings has not been supplied. First-key initialization interruption, power-loss/key-loss recovery, physical accessibility, and reviewed translations also remain outside the proven scope. Agent/source review is not being relabeled as external assurance.
trmosala commentedon Sep 22, 2026
Physical macOS Touch ID acceptance now passed and is published in
e4b4086a76.Using the production
vaultAuthentication.verifyadapter on this Mac:VERIFIED.CANCELLED; the expected-cancellation harness fails if authentication succeeds.This completes the actual Touch ID success/cancel portion alongside the previously recorded Keychain-backed cross-process persistence, capability-loss/corruption/migration/storage-fault, and packaged-helper integrity evidence.
#34 remains open only for assurance that cannot be truthfully supplied by this implementation session: explicit owner/security acceptance of the documented same-user shell-capable-agent boundary, external security review with tracked findings, and the separately documented first-key/power-loss/key-loss recovery limitations. No agent review is being relabeled as external review.
trmosala commentedon Sep 22, 2026
The repository owner explicitly accepted the documented assurance limits for the internal macOS test build on 2026-09-22. The decision is recorded in
7f18709634.Accepted internal-test scope:
Broader distribution still requires the repository's written WPP Open owner sign-off and external security review with tracked findings.
Together with the previously recorded production-adapter Touch ID
VERIFIED/CANCELLEDresults, actual Keychain-backed cross-process persistence, corruption/migration/interruption/fault coverage, and packaged-helper integrity/upgrade evidence, this completes the owner-approved #34 scope.